Automated Detection of Precancerous Lesion and Early Esophageal Squamous Cancer Using a Deep Learning Model

Abstract

Introduction: Early detection of esophageal squamous cancer is essential for favorable prognosis of patients. However the ability to detect precancerous lesion and early esophageal squamous cancer is substantially subject to operator experience. We develop an artificial intelligence model(AIM) for automated detection of these lesions, trying to reduce interobserver variability and improve the diagnosis of early esophageal cancer. Methods: We used 1014 narrow band images(NBI) of precancerous lesions and esophageal squamous cancers in 95 cases and 1071 NBI images of non-cancerous lesions in 198 cases from 2017, to train the model. 1478 malignant NBI images in 59 consecutive cases diagnosed as precancerous lesion or esophageal squamous cancer from January 2018 to February 2018 were included to test the mode. 3899 non-cancerous NBI images in 1205 cases dignosed as normal squamous epithelium(821 cases), esophagitis( 109 cases), heterotopic gastric mucosa(154 cases), esophageal varices(69 cases) and submucosal tumor (52 cases) from 2018 were also tested. AI hot zone image was generated for any input endoscopic image. The red color indicates high possibility of cancerous lesion, while the blue color indicates noncancerous lesion. When AIM detected any precancerous lesion or esophageal squamous cancer, a blue spot was displayed to indicate the lesion of interest. Results: Among 59 cases, 32 cases (total 35 lesions) were diagnosed as precancerous lesion or early esophageal squamous cancer by endoscopic submucosal dissection or surgery. The rest 27 cases(total 29 lesions) were diagnosed as cancerous lesions by biopsy including early and advanced esophageal cancer. The sensitivity of AIM for precancerous lesion and early esophageal squamous cancer images in 32 cases was 97.1%. The sensitivity of AIM for all malignant images in 59 cases was 97.8%. The specificity of AIM for all non-cancerous images was 76.2%. Conclusion: AIM demonstrated high diagnostic accuracy for precancerous lesion and early esophageal cancer. False positive rate was high under current mode. Future effort is warranted to build a more accurate mode.317_A Figure 1. patient characteristic and AI test result317_B Figure 2. AI result of cancerous image317_C Figure 3. AI result of non-cancerous image